fix(amdsmi): stop dropping single-instance xGMI data counters by adam360x · Pull Request #12144 · ROCm/rocm-systems

adam360x · 2026-09-23T20:03:05Z

Motivation

On MI450 the GPU-CPU link is xGMI, and amdgpu publishes that link's read/write traffic in gpu_metrics v1.9 as a single-instance counter. amd-smi skipped single-instance rows when copying into per-link arrays, so xgmi_read_data_acc / xgmi_write_data_acc and the link read / write reported N/A. (The PCIe link fields are N/A by design on MI450: there is no PCIe host link.)

Technical Details

Dynamic metrics mapper (rocm_smi/src/rocm_smi_gpu_metrics.cc, assign_vector):

Store a single-instance row as element 0 of its per-link / per-instance array instead of skipping it
Accept rows of any unsigned width; keep an all-ones "not reported" value, and values that do not fit the field, unset

Tests (tests/amd_smi_test/unit/gpu/dynamic_metrics_test.cc):

Single-instance, per-link, narrower-type, narrower-unset, and out-of-range rows

Changelog (CHANGELOG.md):

Resolved Issues entry for the MI450 xGMI counters

No API or ABI change; wrapper, interface, and CLI are untouched.

Issue Tracking

JIRA ID: ROCM-31364

Test Plan

amdsmitst --gtest_filter='*Unit.*' (hardware-free)
Mutation checks: the previous mapper and a naive cast each fail the new tests
MI455X (gfx1250, gpu_metrics v1.9): amdsmi_get_gpu_metrics_info() and amdsmi_get_link_metrics(), develop vs this branch via AMDSMI_LIB_OVERRIDE

Test Result

158/158 unit tests pass
MI455X: xGMI read/write[0] N/A -> 578484 KB / 0, matching the raw driver table; other array metrics unchanged
amd-smi xgmi still shows read/write as N/A: it lists only peer-GPU hive links, and this system has none (num_links=0)
Separately, xgmi_link_speed reads 32768 on MI450 because the driver copies the Q10 firmware value unrounded; reported to the driver team

Submission Checklist

Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.

therock-pr-bot · 2026-09-23T20:04:00Z

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

therock-pr-bot · 2026-09-23T20:04:01Z

🚫 Please fix the failed policies before requesting reviews.

The following policy checks failed:

❌ PR Description

The Not ready to Review label has been added to this PR.
Once all policies pass, the label will be removed automatically.

Copilot

Copilot review overview

🔵 Needs a closer look

Scalar values with integral types differing from the exact target type are not yet handled.

Review effort: Lite
Findings: None

What changed in this PR

Updates AMD SMI dynamic GPU metrics mapping so single-instance xGMI counters populate link 0.

Changes:

Adds scalar-to-array mapping.
Adds xGMI mapping tests.
Documents the MI450 fix.

File	Summary
`projects/amdsmi/tests/amd_smi_test/unit/gpu/dynamic_metrics_test.cc`	Adds single-instance and per-link xGMI tests.
`projects/amdsmi/rocm_smi/src/rocm_smi_gpu_metrics.cc`	Maps scalar rows to array element 0.
`projects/amdsmi/CHANGELOG.md`	Documents the resolved issue.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

adam360x · 2026-09-23T20:19:24Z

Re: the Copilot review note, "Scalar values with integral types differing from the exact target type are not yet handled". Agreed, addressed in 1f31955.

The single-instance path matched only a scalar of the field's exact type, so a driver that declares the counter narrower (e.g. U32 into a uint64_t field) would still have been dropped.

Single-instance rows of any unsigned width now map to element 0.
A plain static_cast would reintroduce a sentinel bug: a narrower all-ones "not reported" value (e.g. UINT32_MAX) would widen into a real-looking 4294967295, and a wider value would truncate. Both are now kept unset.
Signed rows are still skipped, as before: none of these per-instance fields is signed, and a negative value has no valid unsigned mapping.
New tests: narrower-type, narrower-unset, and out-of-range. The previous mapper fails the first; a naive cast fails the other two (4294967295, 4464).
158/158 unit tests pass. Re-verified on MI455X: xGMI read/write[0] go from N/A to 578484 KB / 0, and the other array metrics are unchanged.

charis-poag-amd

Pre-approving, did have one question though. See inline.

- Store a single-instance dynamic gpu_metrics row as element 0 of its per-link array; the mapper accepted only multi-value rows, so the one xGMI read/write data counter MI450 reports for its GPU-CPU link was dropped and every link read as unavailable - Add unit tests for single-instance and per-link xGMI rows Signed-off-by: Adam360x <Adam.pryor@amd.com>

- Accept a single-instance dynamic gpu_metrics row of any unsigned width, not only the field's exact type, so a driver that declares a narrower counter still fills element 0 - Keep an all-ones "not reported" value, and values that do not fit the field, unset instead of widening or truncating them into plausible readings - Cover narrower, narrower-unset, and out-of-range rows in unit tests Signed-off-by: Adam360x <Adam.pryor@amd.com>

systems-assistant · 2026-10-07T16:51:18Z

TheRock Submodule Bump Activity

Newest first

2026-10-07 — Included in TheRock via ROCm/TheRock#8809.

## Motivation ROCm 10.1 has branched to `release/therock-10.1`. This opens the 10.2 development cycle: it bumps the library minor version to 27.2.0 and fills the `10.2.0` changelog section. While opening it, the `10.1.0` section on `develop` was audited against the release branch, and entries that will not ship in 10.1 were moved to `10.2.0`. ## Technical Details **Version** (`include/amd_smi/amdsmi.h`, `rust-interface/src/amdsmi_wrapper.rs`): - Bump `AMDSMI_LIB_VERSION_MINOR` `1 -> 2` (27.2.0). SONAME stays `libamd_smi.so.27` (`SOVERSION` is `MAJOR` only). - Regenerate the Rust binding constant to match. The Python wrapper and Go shim carry no minor-version constant. **Changelog** (`CHANGELOG.md`): - Extend the existing `## amd_smi_lib for ROCm 10.2.0` section. - Move the `10.1.0` entries whose source PR is not on `release/therock-10.1`, and add entries for three develop-only PRs whose fixes merged without one: | Action | Source PR | Entries | |--------|-----------|---------| | Move `10.1.0` → `10.2.0` | #8624 | CLI exit-code rework: 5 Changed, 1 Removed, 2 Resolved Issues | | Move `10.1.0` → `10.2.0` | #11494 | `AMDSMI_VRAM_TYPE_HBM4` (Added) | | Move `10.1.0` → `10.2.0` | #11066 | fwupd UMA carveout (Changed) | | Move `10.1.0` → `10.2.0` | #11471 | NPM node power limit: 1 Added, 2 Changed | | Move `10.1.0` → `10.2.0` | #10678 | Version-absent `amd-smi metric` fields omitted from human output (Changed; was under Removed, but it changes output) | | Move `10.1.0` → `10.2.0` | #12144 | xGMI data counters reading as unavailable on MI450 (Resolved Issues) | | Move `10.1.0` → `10.2.0` | #12804 | NPM balancing mode get/set (Added) | | New in `10.2.0` | #11474 | Topology APIs returned uninitialized NUMA data (Resolved Issues) | | New in `10.2.0` | #11045 | `amdsmi_get_gpu_fabric_info()` rejected partial UALink sysfs data (Resolved Issues) | | New in `10.2.0` | #12999 | NULL output pointers, CPER paging after a large record, `amd-smi static --clock` traceback (3 Resolved Issues) | - Drop the `### Removed` heading left empty in `10.1.0`. - Keep the #11628 entry in `10.1.0` (cherry-picked to the release branch as #12032) and the `10.0.0` section as is (it differs from `release/therock-10.0` only by the #11046 placement fixes). **Docs** (`docs/how-to/amdsmi-cli-tool.md`): - Update the `amd-smi --help` banner and default-view header examples to 27.2.0 / ROCm 10.2.0 (stale since 27.0.0). ## Issue Tracking N/A ## Test Plan - Configure and build with `-DBUILD_TESTS=ON`; confirm CMake reports 27.2.0. - Run `pre-commit` (amdsmi config) on the changed files. - Check `10.2.0`, `10.1.0` and `10.0.0` against changelog conventions: allowed subsection headings, two-space line breaks on headline bullets, no JIRA IDs in entry text. - Diff the `10.1.0` entries and headings against `release/therock-10.1`. - No GPU run or new unit test: no runtime code changed. ## Test Result - Build: pass (`amdsmitst` links), CMake reports 27.2.0 - `pre-commit`: all hooks passed - Changelog conventions: clean for `10.2.0`, `10.1.0`, `10.0.0` - `10.1.0` vs `release/therock-10.1`: identical entries and headings --------- Signed-off-by: Maisam Arif <Maisam.Arif@amd.com>

adam360x requested review from a team and a lite review from Copilot September 23, 2026 20:03

github-actions Bot added the project: amdsmi label Sep 23, 2026

Copilot started reviewing on behalf of adam360x September 23, 2026 20:03

View session

therock-pr-bot Bot added the Not ready to Review label Sep 23, 2026

Copilot AI reviewed Sep 23, 2026

View reviewed changes

systems-assistant Bot added the organization: ROCm label Sep 23, 2026

oliveiradan approved these changes Sep 23, 2026

View reviewed changes

charis-poag-amd approved these changes Sep 28, 2026

View reviewed changes

Comment thread projects/amdsmi/rocm_smi/src/rocm_smi_gpu_metrics.cc Resolved

adam360x removed the Not ready to Review label Oct 2, 2026

adam360x added 2 commits October 6, 2026 03:39

adam360x force-pushed the users/adam360x/ROCM-31364 branch from 67e19fd to 3bd95f2 Compare October 6, 2026 08:40

marifamd merged commit f1e1a4b into develop Oct 6, 2026
227 of 232 checks passed

marifamd deleted the users/adam360x/ROCM-31364 branch October 6, 2026 15:40

adam360x mentioned this pull request Oct 7, 2026

fix(amdsmi): return GPU metrics on APUs with metrics v2.1-v2.3 #11075

Merged

1 task done

marifamd mentioned this pull request Oct 7, 2026

chore(amdsmi): bump version to 27.2.0 and update CHANGELOG #12623

Merged

1 task done

Navigation Menu

Repository navigation

fix(amdsmi): stop dropping single-instance xGMI data counters - #12144

fix(amdsmi): stop dropping single-instance xGMI data counters#12144
marifamd merged 2 commits into
developfrom
users/adam360x/ROCM-31364

adam360x commented Sep 23, 2026 •

edited

Loading

Uh oh!

therock-pr-bot Bot commented Sep 23, 2026 •

edited

Loading

Uh oh!

therock-pr-bot Bot commented Sep 23, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

adam360x commented Sep 23, 2026

Uh oh!

charis-poag-amd left a comment

Uh oh!

Uh oh!

Uh oh!

systems-assistant Bot commented Oct 7, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

adam360x commented Sep 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Issue Tracking

Test Plan

Test Result

Submission Checklist

Uh oh!

therock-pr-bot Bot commented Sep 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✅ All Policy Checks Passed

Uh oh!

therock-pr-bot Bot commented Sep 23, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🔵 Needs a closer look

Uh oh!

adam360x commented Sep 23, 2026

Uh oh!

charis-poag-amd left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

systems-assistant Bot commented Oct 7, 2026

TheRock Submodule Bump Activity

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

adam360x commented Sep 23, 2026 •

edited

Loading

therock-pr-bot Bot commented Sep 23, 2026 •

edited

Loading